Video-based face tracking and recognition on updating twin GMMs

Online learning is a very desirable capability for video-based algorithms. In this paper, we propose a novel framework to solve the problems of video-based face tracking and recognition by online updating twin GMMs. At first, considering differences between the tasks of face tracking and face recognition, the twin GMMs are initialized with different rules for tracking and recognition purposes, respectively. Then, given training sequences for learning, both of them are updated with some online incremental learning algorithm, so the tracking performance is improved and the class-specific GMMs are obtained. Lastly, Bayesian inference is incorporated into the recognition framework to accumulate the temporal information in video. Experiments have demonstrated that the algorithm can achieve better performance than some well-known methods.